Learning to Predict Grasp Reliability for a Multifinger Robot Hand by Using Visual Features

A. Morales, E. Chinelalto, P.J. Sanz, A.P. del Pobil (Spain), and A.H. Fagg (USA)


Robot grasping, dextereous manipulation, active learning.


This paper describes a practical approach to the robot grasping problem. An approach that is composed of two different parts. First, a vision-based grasp synthesis sys tem implemented on a humanoid robot able to compute a set of feasible grasps and to execute any of them. This grasping system takes into account gripper kinematics con straints and uses little computational effort. Second, a learning framework aimed at discovering the visual features that predict a reliable grasp. A grasp characterization scheme based on a set of visual features is developed in order to describe and compare grasps. In ad dition, a practical measure of grasp reliability is designed and implemented.Moreover, an algorithm aimed at predict ing the performance of an untested grasp using the results observed on previous similar attempts is presented. A sec ond algorithm that actively selects the next grasp to be ex ecuted in order to improve the predictive quality of the ac cumulated experience is introduced, too. An exhaustive database of experimental data is col lected and used to test and validate both algorithms.

Important Links:

Go Back